Mónica G. Larese

Automatic classification of legumes using leaf vein image features

In this paper, a procedure for segmenting and classifying scanned legume leaves based only on the analysis of their veins is proposed (leaf shape, size, texture and color are discarded). Three legume species are studied, namely soybean, red and white beans. The leaf images are acquired using a standard scanner. The segmentation is performed using the unconstrained hit-or-miss transform and adaptive thresholding. Several morphological features are computed on the segmented venation, and classified using four alternative classifiers, namely support vector machines (linear and Gaussian kernels), penalized discriminant analysis and random forests. The performance is compared to the one obtained with cleared leaves images, which require a more expensive, time consuming and delicate procedure of acquisition. The results are encouraging, showing that the proposed approach is an effective and more economic alternative solution which outperforms the manual experts recognition. HighlightsWe develop an automatic procedure to classify legume species using scanned leaves.The method is based exclusively on the analysis of the leaf venation images.We analyze the advantages over the usage of cleared leaves.Different state-of-the-art classifiers are compared.The proposed method outperforms human expert classification.

Deep learning for plant identification using vein morphological patterns

Display Omitted Deep convolutional neural network (CNN) for plant identification focusing on leaf vein patterns.No task-specific feature extractors needed.Improved the state of the art accuracy on a legume species recognition task.Visualization of relevant vein patterns. We propose using a deep convolutional neural network (CNN) for the problem of plant identification from leaf vein patterns. In particular, we consider classifying three different legume species: white bean, red bean and soybean. The introduction of a CNN avoids the use of handcrafted feature extractors as it is standard in state of the art pipeline. Furthermore, this deep learning approach significantly improves the accuracy of the referred pipeline. We also show that the reported accuracy is reached by increasing the model depth. Finally, by analyzing the resulting models with a simple visualization technique, we are able to unveil relevant vein patterns.

Multiscale recognition of legume varieties based on leaf venation images

Abstract In this work we propose an automatic low cost procedure aimed at classifying legume species and varieties based exclusively on the characterization and analysis of the leaf venation network. The identification of leaf venation patterns which are characteristic for each species or variety is not an easy task since in some situations (specially for cultivars from the same species) the vein differences are visually indistinguishable for humans. The proposed procedure takes as input leaf images acquired using a standard scanner, processes the images in order to segment the veins at different scales, and measures different traits on them. We use these features in combination with modern automatic classifiers and feature selection techniques in order to perform recognition. The process was initially applied to recognize three different legumes in order to evaluate the improvements over previous works in the literature, and then it was employed to distinguish three diverse soybean cultivars. The results show the improvements achieved by the usage of the multiscale features. The cultivar recognition is a more challenging problem, since the experts cannot distinguish evident differences in plain sight. However, we achieve acceptable classification results. We also analyze the feature relevance and identify, for each classifier, a small set of distinctive traits to differentiate the species and varieties.

Gene set enrichment analysis using non-parametric scores

Gene Set Enrichment Analysis (GSEA) is a well-known technique used for studying groups of functionally related genes and their correlation with phenotype. This method creates a ranked list of genes, which is used to calculate an enrichment score. In this work, we introduce two different metrics for gene ranking in GSEA, namely the Wilcoxon and the Baumgartner-Weis-Schindler tests. The advantage of these metrics is that they do not assume any particular distribution on the data. We compared them with the signal-to-noise ratio metric originally proposed by the developers of GSEA on a type 2 diabetes mellitus (DM2) database. Statistical significance is evaluated by means of false discovery rate and p-value calculations. Results show that the Baumgartner-WeisSchindler test detects more pathways with statistical significance. One of them could be related to DM2, according to the literature, but further research is needed.

Spot defects detection in cDNA microarray images

Bad quality spots should be filtered out at early steps in microarray analysis to avoid noisy data. In this paper we implement quality control of individual spots from real microarray images. First of all, we consider the binary classification problem of detecting bad quality spots. We propose the use of ensemble algorithms to perform detection and obtain improved accuracies over previous studies in the literature. Next, we analyze the untackled problem of identifying specific spot defects. One spot may have several faults simultaneously (or none of them) yielding a multi-label classification problem. We propose several extra features in addition to those used for binary classification, and we use three different methods to perform the classification task: five independent binary classifiers, the recent Convex Multi-task Feature Learning (CMFL) algorithm and Convex Multi-task Independent Learning (CMIL). We analyze the Hamming loss and areas under the receiver operating characteristic (ROC) curves to quantify the accuracies of the methods. We find that the three strategies achieve similar results leading to a successful identification of particular defects. Also, using a Random forest-based analysis we show that the newly introduced features are highly relevant for this task.

Finding local leaf vein patterns for legume characterization and classification

In recent years, the importance of analyzing the effect of genetic variations on the plant phenotypes has raised much attention. In this paper, we describe a procedure which can be useful to discover representative leaf vein patterns for each species or variety under analysis. We consider three legumes, namely red bean, white bean and soybean. Soybean specimens are also divided in three cultivars. In total there are five leaf vein image classes. In order to find the discriminative patterns, we detect Self-Invariant Feature Transform (SIFT) keypoints in the segmented vein images. The Bag of Words model is built using SIFT descriptors, and classification is performed resorting to Support Vector Machines with a Gaussian kernel. Classification accuracies outperform recent results available in the literature and manual classification, showing the advantages of the procedure. The Bag of Words model is useful for vein patterns characterization and provides a means to highlight the most representative patterns for each species and variety.

Clustering stability for automated color image segmentation

Abstract Clustering is a well-established technique for segmentation. However, clustering validation is rarely used for this purpose. In this work we adapt a clustering validation method, Clustering Stability (CS), to automatically segment images. CS is not limited by image dimensionality nor by the clustering algorithm. We show clustering and validation acting together as a data-driven process able to find the optimum number of partitions according to our proposed color-texture feature representation. We also describe how to adapt CS to detect the best settings required for feature extraction. The segmentation solutions found by our method are supported by a stability score named STI, which provides an objective quantifiable metric to obtain the final segmentation results. Furthermore, the STI allows to compare multiple alternative solutions and select the most appropriate according to the index meaning. We successfully test our procedure on texture and natural images, and 3D MRI data.

Hybrid Consensus Learning for Legume Species and Cultivars Classification

In this work we propose an automatic method aimed at classifying five legume species and varieties using leaf venation features. Firstly, we segment the leaf veins and measure several multiscale morphological features on the vein segments and the areoles. Next, we build a hybrid consensus of experts formed by five different automatic classifiers to perform the classification using the extracted features. We propose to use two strategies in order to assign the importance to the votes of the algorithms in the consensus. The first one is considering all the algorithms equally important. The second one is based on the accuracy of the standalone classifiers. The performance of both consensus classifiers show to outperform the standalone classification algorithms in the five class recognition task.

Seed-per-pod estimation for plant breeding using deep learning

Abstract Commercial and scientific plant breeding programs require the phenotyping of large populations. Phenotyping is typically a manual task (costly, time-consuming and sometimes arbitrary). The use of computer vision techniques is a potential solution to some of these specific tasks. In the last years, Deep Learning, and in particular Convolutional Neural Networks (CNNs), have shown a number of advantages over traditional methods in the area. In this work we introduce a computer vision method that estimates the number of seeds into soybean pods, a difficult task that usually requires the intervention of human experts. To this end we developed a classic approach, based on tailored features extraction (FE) followed by a Support Vector Machines (SVM) classification model, and also the referred CNNs. We show how standard CNNs can be easily configured and how a simple method can be used to visualize the key features learned by the model in order to infer the correct class. We processed different seasons batches with both methods obtaining 50.4% (FE + SVM) and 86.2% (CNN) of accuracy in test, highlighting the particularly high increase in generalization capabilities of a deep learning approach over a classic machine vision approach in this task. Dataset and code are publicly available.

Quantitative Improvements in cDNA Microarray Spot Segmentation

When developing a cDNA microarray experiment, the segmentation of individual spots is a crucial stage. Spot intensity measurements and gene expression ratios directly depend on the effectiveness and accuracy of the segmentation results. However, since the ground truth is unknown in microarray experiments, quantification of the accuracy of the segmentation process is a very difficult task. In this paper an improved unsupervised technique based on the combination of clustering algorithms and Markov Random Fields (MRF) is proposed to separate the foreground and background intensity signals used in the spot ratio computation. The segmentation algorithm uses one of two alternative methods to provide for initialization, namely K-means and Gaussian Mixture Models (GMM) clustering. This initial segmentation is then processed via MRF. Accuracy is measured by means of a set of microarray images containing spike spots where the target ratios are known a priori , thus making it possible to quantify the expression ratio errors. Results show improvements over state-of-the-art procedures.